Loading...

Design and Implementation of an Intelligent Control System Based-on Deep Reinforcement Learning for a Lower-limb Hybrid Exoskeleton Robot

Koushki, Amir Reza | 2021

243 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 54830 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Vossoughi, Gholamreza; Boroushaki, Mehrdad
  7. Abstract:
  8. Hybrid Exoskeletons refer to simultaneous use of wearable robots and functional electrical stimulation technology. Hybrid exoskeletons have many advantages compared to the separate application of each of these technologies, such as reducing the robot’s energy consumption and the need for lighter and cheaper actuators for the robot, using humans muscle power, and reducing muscle fatigue. As a result, these robots have recently attracted a lot of interest in rehabilitation applications for patients suffering from mobility impairment.Control in hybrid exoskeletons is more complicated than control in traditional exoskeletons. Because in addition to robot and functional electrical stimulation control, a strategy should also be designed to coordinate or allocate between these two parallel systems. The subject of allocation in these robots is a significant challenge due to the complexities of system modeling, which make online and dynamic optimization of this parameter difficult. As a result, many previous studies have avoided considering the allocation between the robot and FES or have done so by simplifying models and optimizing single-objectives such as fatigue alone. To address this issue, in this research, a novel model-free control strategy based on deep reinforcement learning algorithms has been developed for hybrid wearable robots for rehabilitation purposes. In this control strategy, for the first time, a DQN supervisor is used to obtain online, dynamic, and intelligent allocation between the robot and FES. This supervisor has the task of performing multi-objective optimization consisting of trajectory tracking error, robot power consumption, and user muscle fatigue.This strategy was first implemented on a simulation platform of hybrid exoskeletons. After success and proof in this stage, it was implemented on the experimental setup of the Sharif Hybrid Exoskeleton in the oscillating phase of the knee joint gate. After training the DQN networks for 200 episodes on the experimental robot platform, the results obtained from the implementation of the control strategy, showed that the developed approach was able to track the desired trajectory with an RMS tracking error of 1.1 degrees and an average fatigue factor of 0.86 while reducing the average energy consumption of the robot by about 24.5% compared to the exoskeleton without FES. These results showed that the proposed control strategy based on the DQN supervisor has been able to successfully perform the control and optimization of the mentioned parameters. The strategy has succeeded in tracking the desired trajectory with a small error while reducing robot energy consumption. At the same time, it has been able to manage fatigue and prevent fatigue caused by functional electrical stimulation in the user’s muscles. Finally, the results were compared with previous studies in this field and proved the superiority of the proposed method
  9. Keywords:
  10. Functional Electrical Stimulation ; Rehabilitation ; Deep Reinforcement Learning ; Muscular Fatigue ; Exoskeleton ; Lower Limb Exoskeleton ; Deep Q-Network (DQN)Algorithm ; Intelligent Control

 Digital Object List

 Bookmark

No TOC